A genetic aggregate stereo algorithm for 3-D classification of occluded shapes

摘要

In computer vision, various stereo algorithms have been developed to determine the shape depth from a pair of images for that shape. To find the depth traditionally, one has to solve the point-to-point correspondence problem. By nature, this problem is complicated and time consuming. In this paper, a genetic algorithm has been exploited to obtain a depth estimation for the underlying shape without using any point-to-point correspondence. Instead, correspondence of the outer contours as a whole is required. Consequently, the three-dimensional (3-D) moments as well as a chain code, could be computed and relied upon as shape invariant features. Such features are used as input to a next stage that consists of a neural network classifier. Combining the robustness of neural network classifiers with the robustness of genetic algorithms has led to a fairly robust pattern recognition system that can tolerate high degrees of noise and occlusion levels. Eventually, a dynamic alignment procedure is attached to the classifier outputs to make the proposed shape representation independent of the starting point.